Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model
The paper says the failure is in the training signal: the model can keep its frozen representation, but its flow has to be taught through decoded rollouts to recover motion.
The authors report that a pretrained flow leaves manipulated objects still, while latent-only retraining produces abrupt, teleport-like movement. Their DART method retrains only the flow, adding decode-path supervision so changes land in the right part of the rollout. They say it restores temporal motion structure, reconnects predicted motion to the scene, and closes nearly half the remaining gap to an oracle-informed interpolation reference at larger scale. They also warn that pixel error by itself can favor frozen predictions. ArXiv · AI/CL/LG's note
The authors report that a pretrained flow leaves manipulated objects still, while latent-only retraining produces abrupt, teleport-like movement. Their DART method retrains only the flow, adding decode-path supervision so changes land in the right part of the rollout. They say it restores temporal motion structure, reconnects predicted motion to the scene, and closes nearly half the remaining gap to an oracle-informed interpolation reference at larger scale. They also warn that pixel error by itself can favor frozen predictions. ArXiv · AI/CL/LG's note
score 5